arXiv Computer Vision

Latent Spatial Memory for Video World Models

The paper introduces latent spatial memory, a 3D cache that stores scene information directly in diffusion latent space, eliminating the need for pixel-space reconstruction. It presents Mirage, a framework that lifts latent tokens into 3D using depth-guided back‑projection and queries the memory via latent‑space warping, achieving significant speed and memory gains. Experiments demonstrate up to 10.57× faster video generation, a 55× reduction in memory usage, and state‑of‑the‑art performance on WorldScore and strong reconstruction on RealEstate10K.

Hugging Face Trending Papers
Jun 8

Latent Spatial Memory for Video World Models

Video world models that maintain 3D spatial consistency across generated frames typically rely on explicit point cloud memory constructed in RGB space. This design is both computationally expensive, requiring repeated rendering and VAE encoding, and inherently lossy, as the round trip through pixel space discards rich features of the learned latent representation.

arXiv Computer Vision
Oct 19

Persistent Planar Memory for Video World Models

Persistent Planar Memory for Video World Models (PPMem) introduces a compact 3D memory system using plane primitives for video world models. By integrating a planar prediction head and Depth Token Fusion into a video diffusion transformer, PPMem jointly generates RGB video and planar geometry, merging them into a persistent plane memory that conditions future video chunks. The resulting memory is highly compact, geometrically explicit, cross-view consistent, and adaptive, outperforming prior 3D-grounded models on DL3DV and RealEstate10K in video quality and depth/normal accuracy.

By Yuze He, Bin Tan, Zelin Gao, Yujun Shen, Yong-Jin Liu, Nan Xue, Yinghao Xu
arXiv Computer Vision
Nov 2

PRISM: Feed-Forward Single-Image 3D Reconstruction via Geometric Warp-Residual Modeling

PRISM is a feed‑forward single‑image 3D reconstruction framework that splits the task into a parameter‑free geometric warp and a learned residual correction, eliminating the need for iterative diffusion sampling. It uses a two‑stage training strategy—latent supervised distillation for geometric generalization and perceptual fine‑tuning for appearance quality—to generalize from synthetic data. Experiments on three benchmarks show PRISM matches diffusion‑based methods in quality while cutting inference time to just 36 seconds per scene.

By Zhijie Zheng, Xinhao Xiang, Jiawei Zhang
arXiv Computer Vision
Oct 21

Latent Bridge: From Video Latents to 3D Geometry via Feature Alignment

Latent Bridge connects video generative latent spaces with geometric feature spaces, enabling direct prediction of 3D geometry—cameras, depth, and point clouds—without an intermediate RGB decoding step. The method trains representation‑specific bridges for Wan2.1 and V‑RAE via feature distillation on RGB videos while keeping the video encoder and geometry model frozen, requiring no ground‑truth geometry. At inference, the same video latents support both RGB decoding and geometry prediction, and the trained bridge can be reused across image‑to‑video and text‑to‑video generation, achieving competitive reconstruction quality on Co3Dv2 and WildRGB‑D with efficient end‑to‑end generation.

By Shuai Zhang, Zhanpeng Ouyang, Zhengqing Chen, Ruohong Mei, Shaoyu Chen, Weiqiang Ren, Hui Sun, Qian Zhang, Wenyu Liu, Xinggang Wang
arXiv Computer Vision
Aug 28

SpatialCrafter: Single Image World Modeling with Generative 3D Proxies

SpatialCrafter introduces a two‑stage framework for single‑image world modeling that first generates a global 3D proxy using a Point‑anchored Sparse Structure Flow module, then refines appearance with a Generative Deferred Refiner built on a video diffusion model. The method incorporates Parallel Geometry Injection and Proxy‑Aware Corruption training to integrate the proxy without disrupting the pretrained generative manifold, and it is evaluated on a newly constructed dataset of 115K scenes. Experiments demonstrate that SpatialCrafter outperforms existing approaches, reducing long‑term drift and maintaining consistency under rapid camera motion and extreme viewpoints.

By Chuan Fang, Lingteng Qiu, Yixun Liang, Rui Chen, Kunming Luo, Zhaohua Zheng, Tongyuan Bai, Feipeng Tian, Zilong Dong, Zihan Zhou, Ping Tan
arXiv Computer Vision
Sep 4

OctWorld: Long-Range World-Consistent Video Generation with Octree-Based 3D Mapping

OctWorld is a video diffusion framework that uses a persistent 3D memory called OctMap to generate long‑range, world‑consistent videos from a single image. OctMap fuses visual observations and depth maps into a dynamic sparse octree, adapting spatial resolution to image evidence while keeping memory usage low. Experiments show that OctWorld produces spatially consistent videos over extended camera paths and outperforms previous methods on both standard benchmarks and challenging long‑range scenarios.

By Zelong Lv, Sicheng Xu, Jianfeng Xiang, Ruicheng Wang, Yue Dong, Yu Deng, Guangzhong Sun, Jiaolong Yang